An Efficient Seizure Prediction System using Convolutional Neural Network and Scalogram of the EEG Signals

Epilepsy is a life-threatening neurological illness that affects millions of people throughout the world. The most common symptom of epilepsy is seizures. Seizures are characterized by fast bursts of uncontrolled electrical activity in a group of neurons in specific areas of the cerebral cortex. The method used most frequently in clinical practice to identify epileptic seizures is electroencephalography (EEG). But manual examination of EEG brain waves is a laborious and error-prone operation that takes time. A number of automatic systems have already been developed to help the neurologists to identify epilepsy. But the aberrant activity related to the seizure starts a few minutes before the onset of seizure. As a result, if the system can detect discriminative traits from the pre-ictal time period, the seizure prediction procedure can be automated as well. Fast and reliable seizure prediction is still a difficult task due to the non-stationary nature of EEG signal. The proposed work introduces an efficient deep learning approach for seizure anticipation in which the time series EEG signal is first converted to a two-dimensional color image called scalogram and fed to the neural network for feature extraction. CNN model is used for extracting the features. The NN is employed with separable convolution to offer the best prediction result in epileptic seizures. Multiple units of CNNs which operates on the same scalogram are connected in parallel. The outputs of the parallel neural networks with different discriminative features are concatenated so that classification power can be improved. A dense layer with an output layer is used for classification. The proposed approach was tested on 22 human beings from the CHBMIT scalp EEG dataset, and it was found that the suggested model provides an accuracy of 98.3%, specificity of 100%, and sensitivity is 96.68%.